$9M Seed Backs Probably's Deterministic AI Validation Layer

San Francisco startup Probably has secured $9 million in seed funding co-led by Andreessen Horowitz and Accel to develop a deterministic validation layer for local AI inference. Founded by Peter Elias, the company utilizes a framework designed to force smaller language models to achieve 99.99% accuracy on structured data tasks. The funding round, which included participation from Tokyo Black and Vermilion Cliffs Ventures, signals a shift toward strict data auditability over generalized model scaling.

Deterministic Validation Architecture

Probably relies on a framework Elias refers to as a “data science mech suit.” The system wraps target models in a harness that strictly checks generated answers against verifiable raw data. If the validation layer detects any conflict between the generated text and the underlying dataset, the response is immediately rejected and regenerated.

This separation of concerns allows developers to address why AI hallucinates by moving verification outside the probabilistic weights of the model itself. The framework processes data entirely on local hardware, currently optimized for Apple Silicon M1 through M5 processors. By utilizing DuckDB for analytical processing, the tool guarantees that sensitive information never leaves the local machine or private network. Every output includes explicit citations and a continuous audit trail to meet compliance requirements in finance, legal, and healthcare applications.

Model Tiering and Benchmark Targets

Rather than relying on frontier capabilities, Probably explicitly targets models operating four classes below state-of-the-art systems like GPT-5 or Claude 4. The company utilizes these smaller models strictly for natural language translation, delegating computational tasks to traditional deterministic engines.

Architectural Component	Standard LLM Deployment	Probably Validation Harness
Primary Compute Location	Cloud infrastructure	Local Apple M1-M5 via DuckDB
Target Model Tier	Frontier (GPT-5 class)	Four classes weaker
Error Rate Target	Variable by prompt	99.99% (1 error per 10,000)
Mathematical Execution	Probabilistic generation	Delegated local compute engine

This structure limits infrastructure overhead while drastically reducing LLM API costs. The tradeoff relies entirely on the quality of the harness engineering to maintain the 99.99% accuracy target on precision-sensitive queries.

Verifiable Data Agent Beta

The company’s first commercial release is the Verifiable Data Agent, currently in Beta 0.1. The tool enables users to execute natural language queries against complex local and remote datasets. To maintain data privacy, the model interfaces solely with dataset metadata and summary statistics.

The beta supports local CSV, JSON, and Parquet file formats. It also connects directly to enterprise data warehouses including Snowflake, BigQuery, and Postgres. When a query requires calculation, the agent bypasses the language model entirely and routes the mathematical operations to the processor-optimized compute engine.

Recent enterprise surveys from Bain indicate that 40% of companies report AI cost savings of less than 10%. This data reflects a broader industry shift from experimental AI deployments to a period of strict revenue accountability. Enterprises require systems that eliminate compounding token bills while guaranteeing the deterministic accuracy necessary for production environments.

If you build data-sensitive multi-agent systems, decoupling natural language routing from mathematical execution provides a secure path to consistent enterprise analytics.